parameter suggestion for non-neural pixel probes

[chiyu1203]

Dear UnitMatch team,
Thanks for putting this library together.
I do acute recording in walking insects with 64- or 32-channels silicone probe.
In my experiment, I split an animal's recording into consecutive 1-hour sessions within the same day so I can optimise the recording condition. With this library, I was hoping to give a quantitative measurement about to what extent units identified across sessions are in fact the same unit. I have gone through the Demo notebooks in UnitMatchPy and was able to get some results.
I applied UnitMatchPy with default parameters to two animals.

The first animal's result is below

The percentage of units matched to themselves is: 66.67%
The percentage of false -ve's then is: 33.33%
The rate of miss-match(es) per expected match 0.28
The percentage of false +ve's is 0.59% for session 1
The percentage of false +ve's is 0.92% for session 2
The percentage of false +ve's is 1.09% for session 3

The second animal's result is below

The percentage of units matched to themselves is: 66.67%
The percentage of false -ve's then is: 33.33%
The rate of miss-match(es) per expected match 0.28
The percentage of false +ve's is 0.59% for session 1
The percentage of false +ve's is 0.92% for session 2
The percentage of false +ve's is 1.09% for session 3

I actually do not 66.67% is high or low so I checked individual units on the GUI.
It seems that only UM probability over 90% is much more convincing, but even at that level, I saw this kind of matches.

In the end, I am not sure how to evaluate the performance of UnitMatch for my dataset amd whether I should keep using it or switch to DeepUnitMatch. Could anyone suggest me how to change parameters of UnitMatch to improve the performance?
Note1: the density of channels on these probes (32 channels/shank, spanning around 300 or 400 um vertically) is similar or a little bit less than neural pixels 1.0.
Note2: I used kilosort4 to sort out putatitive spikes. In the first animal, I manually labeled Good units, in the second animal, I did not.
Note3: The preprocessed pipeline in the demo notebooks (extract_raw_data_demo_open_ephys.ipynb and UMPy_example.ipynb) were used. And I have not used Bombcell yet.
UnitMatchPy version: 3.3